daveadmin/dobetternorge-tools
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fix legal analysis issue extraction for long documents

Browse Source
This commit is contained in:
davegilligan
2026-05-25 19:06:23 +02:00
parent 190f639784
commit 8205a22205
3 changed files with 263 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
includes/LegalAnalysisAgent.php
+234 -1
View File
@@ -40,7 +40,7 @@ final class DbnLegalAnalysisAgent
*
* @return array<int,array{id:int,question:string,brief_context:string,doc_type:string,severity_hint:string}>
*/
public function extractIssues(string $text, string $language, string $docType): array
private function extractIssuesFromSingleChunk(string $text, string $language, string $docType): array
{
$locale = dbnToolsLanguageName($language);
$text = mb_substr($text, 0, 24000, 'UTF-8'); // keep prompt within 4o-mini context
@@ -118,6 +118,239 @@ PROMPT;
return $clean;
}
/**
* Pass 1 - extract distinct legal issues from representative document windows.
*
* @return array<int,array{id:int,question:string,brief_context:string,doc_type:string,severity_hint:string}>
*/
public function extractIssues(string $text, string $language, string $docType): array
{
$text = trim($text);
if ($text === '') {
return [];
}
$allIssues = [];
foreach ($this->issueExtractionChunks($text) as $chunk) {
try {
$chunkIssues = $this->extractIssuesFromSingleChunk((string)$chunk['text'], $language, $docType);
} catch (Throwable $e) {
error_log('legal-analysis issue extraction failed for ' . (string)$chunk['label'] . ': ' . $e->getMessage());
$chunkIssues = [];
}
foreach ($chunkIssues as $issue) {
$this->appendUniqueIssue($allIssues, $issue);
if (count($allIssues) >= self::MAX_ISSUES) {
break 2;
}
}
}
if (!$allIssues && $this->looksLikeSubstantiveFamilyLawDocument($text, $docType)) {
$allIssues = $this->fallbackLegalIssues($language, $docType, $text);
}
foreach ($allIssues as $idx => &$issue) {
$issue['id'] = $idx + 1;
}
unset($issue);
return array_slice($allIssues, 0, self::MAX_ISSUES);
}
/**
* @return array<int,array{label:string,text:string,offset:int,score:int}>
*/
private function issueExtractionChunks(string $text): array
{
$len = mb_strlen($text, 'UTF-8');
$window = 24000;
if ($len <= $window + 4000) {
return [[
'label' => 'full document',
'text' => $text,
'offset' => 0,
'score' => 0,
]];
}
$chunks = [];
$add = function (string $label, int $offset, int $score = 0) use (&$chunks, $text, $len, $window): void {
$offset = max(0, min($offset, max(0, $len - $window)));
foreach ($chunks as $existing) {
if (abs((int)$existing['offset'] - $offset) < 6000) {
return;
}
}
$chunks[] = [
'label' => $label,
'text' => mb_substr($text, $offset, $window, 'UTF-8'),
'offset' => $offset,
'score' => $score,
];
};
$add('beginning of document', 0, 1);
$add('middle of document', (int)floor(($len - $window) / 2), 1);
$add('end of document', $len - $window, 1);
$keywords = [
'samvaer', 'samvær', 'omsorg', 'barnevern', 'sakkyndig', 'risiko',
'tilknytning', 'rus', 'vold', 'emk', 'barnets beste', 'foreldre',
'bekymring', 'kontakt', 'plassering', 'fylkesnemnd', 'retten',
];
$candidates = [];
for ($offset = 0; $offset < $len; $offset += 10000) {
$chunk = mb_substr($text, $offset, $window, 'UTF-8');
if ($chunk === '') {
break;
}
$lower = mb_strtolower($chunk, 'UTF-8');
$score = 0;
foreach ($keywords as $kw) {
$score += substr_count($lower, mb_strtolower($kw, 'UTF-8'));
}
if ($score > 0) {
$candidates[] = ['offset' => $offset, 'score' => $score];
}
if ($offset + $window >= $len) {
break;
}
}
usort($candidates, static fn(array $a, array $b): int => ($b['score'] <=> $a['score']));
foreach (array_slice($candidates, 0, 4) as $candidate) {
$add('keyword-heavy legal section', (int)$candidate['offset'], (int)$candidate['score']);
if (count($chunks) >= 6) {
break;
}
}
usort($chunks, static function (array $a, array $b): int {
if ($a['score'] !== $b['score']) {
return $b['score'] <=> $a['score'];
}
return $a['offset'] <=> $b['offset'];
});
return array_slice($chunks, 0, 6);
}
/**
* @param array<int,array{id:int,question:string,brief_context:string,doc_type:string,severity_hint:string}> $issues
* @param array{id:int,question:string,brief_context:string,doc_type:string,severity_hint:string} $candidate
*/
private function appendUniqueIssue(array &$issues, array $candidate): void
{
$candidateKey = $this->issueDedupeKey((string)$candidate['question']);
foreach ($issues as $existing) {
$existingKey = $this->issueDedupeKey((string)$existing['question']);
if ($candidateKey === $existingKey || $this->issueSimilarity($candidateKey, $existingKey) >= 0.58) {
return;
}
}
$issues[] = $candidate;
}
private function issueDedupeKey(string $question): string
{
$question = mb_strtolower($question, 'UTF-8');
$question = preg_replace('/[^\p{L}\p{N}\s]+/u', ' ', $question) ?? $question;
$question = preg_replace('/\s+/u', ' ', trim($question)) ?? $question;
return $question;
}
private function issueSimilarity(string $a, string $b): float
{
$aWords = array_values(array_unique(array_filter(preg_split('/\s+/u', $a) ?: [], static fn($w) => mb_strlen($w, 'UTF-8') > 3)));
$bWords = array_values(array_unique(array_filter(preg_split('/\s+/u', $b) ?: [], static fn($w) => mb_strlen($w, 'UTF-8') > 3)));
if (!$aWords || !$bWords) {
return 0.0;
}
$intersection = count(array_intersect($aWords, $bWords));
$union = count(array_unique(array_merge($aWords, $bWords)));
return $union > 0 ? $intersection / $union : 0.0;
}
private function looksLikeSubstantiveFamilyLawDocument(string $text, string $docType): bool
{
if (mb_strlen($text, 'UTF-8') < 8000) {
return false;
}
if (in_array($docType, ['barnevernet', 'adopsjon', 'emergency', 'samvær', 'fylkesnemnd'], true)) {
return true;
}
$lower = mb_strtolower($text, 'UTF-8');
$hits = 0;
foreach (['sakkyndig', 'barnevern', 'barnets beste', 'samvær', 'samvaer', 'omsorg', 'tilknytning', 'emk', 'fylkesnemnd'] as $kw) {
if (str_contains($lower, mb_strtolower($kw, 'UTF-8'))) {
$hits++;
}
}
return $hits >= 2;
}
/**
* @return array<int,array{id:int,question:string,brief_context:string,doc_type:string,severity_hint:string}>
*/
private function fallbackLegalIssues(string $language, string $docType, string $text): array
{
$context = mb_substr(preg_replace('/\s+/u', ' ', trim($text)) ?? trim($text), 0, 300, 'UTF-8');
if ($language === 'no') {
return [
[
'id' => 1,
'question' => 'Hvordan skal barnets beste og samvaer vurderes etter norsk rett?',
'brief_context' => 'Langt familie- eller barnevernsdokument der modellen ikke identifiserte strukturerte spørsmål. Utdrag: ' . $context,
'doc_type' => $docType,
'severity_hint' => 'high',
],
[
'id' => 2,
'question' => 'Er den sakkyndige vurderingen og bevisgrunnlaget tilstrekkelig for konklusjonene?',
'brief_context' => 'Dokumentet ser ut til å inneholde sakkyndige eller faktiske vurderinger som bør testes juridisk.',
'doc_type' => $docType,
'severity_hint' => 'medium',
],
[
'id' => 3,
'question' => 'Er saksbehandling, kontradiksjon og offentlige plikter oppfylt etter norsk rett og EMK?',
'brief_context' => 'Lang sak bør vurderes for prosessuelle rettigheter, dokumentasjonsplikt og forholdsmessighet.',
'doc_type' => $docType,
'severity_hint' => 'medium',
],
];
}
return [
[
'id' => 1,
'question' => 'How should the child best-interests and contact/visitation assessment be reviewed under Norwegian law?',
'brief_context' => 'Long family-law or child-welfare document where the model did not return structured issues. Excerpt: ' . $context,
'doc_type' => $docType,
'severity_hint' => 'high',
],
[
'id' => 2,
'question' => 'Is the expert assessment and evidentiary basis sufficient for the conclusions reached?',
'brief_context' => 'The document appears to contain expert or factual assessments that require legal testing.',
'doc_type' => $docType,
'severity_hint' => 'medium',
],
[
'id' => 3,
'question' => 'Were procedural fairness, contradiction rights, and public-authority duties satisfied under Norwegian law and ECHR?',
'brief_context' => 'A long case file should be checked for procedural rights, documentation duties, and proportionality.',
'doc_type' => $docType,
'severity_hint' => 'medium',
],
];
}
/**
* Pass 2 — single targeted question to dbn-legal-agent-v3 with corpus context.
* Ocelot-only. Capped at 350 tokens / 60s to avoid the documented loop bug.